The key task of advertising media planners is to determine the best media schedule of advertising exposures for a certain budget.The planner could choose the media schedule in one of two ways —a continuous strategy or a pulsing strategy. In a continuous strategy the planner schedules advertising more or less even lyover all the weeks, so as to be present at all times. In a pulsingstrategy the planner advertises in some weeks, often heavily, andnot at all in other times. In this pulsing option the media plannershopes that the heavy advertising conducted in one period will beremembered favorably over time, especially at times when makingpurchase decisions.
This paper offers a methodology that simulates the salesgenerated through the year using different scheduling strategiesand chooses the one which maximizes sales.
We obtained 30 weeks of data from an FMCG marketer on amonthly basis. The data is as follows:
The first stage in the process is building a robust market mix model. This regression model is essentially an equation that seeks to establish the relationship between sales and the input variables.
Sales = F( Base Sales A + Distribution - Price + TV ad stock +Magazine + Outdoor - Competition TV + Seasonal Effect + error (1)TV here enters as ad-stock and not as spends or GRPs.
This way we incorporate both the carry over effects of advertising and its non linear impact on sales where diminishing returns set in.
We used Vector Auto Regression technique to obtain the parameter estimates, but one can also use OLS or MLE to estimate the parameters. All the parameters were significant individually using the't' tests. The model was also empirically validates using the Mean Absolute Percentage Error criteria. This model also had business acceptance.
In the second step we use Non Linear Programming (NLP)technique to obtain a GRP schedule that maximizes sales. The market mix equation is treated the objective function in the NLP model. This is exactly the reverse of the regression model. In the regression we provided the GRP every month as ad stock and the regression provided the parameters. In NLP we provide the Beta parameters and the outputs the GRPs. Business constraints include the fact that the overall GRP cannot be exceeded and that there are thresholds and caps on monthly GRP to be deployed.
Using the estimation method of Congruent Gradient in solving the NLPproblem, we obtained a several schedules that maximized the sales.These sales we higher that the obtained sales, week on week. Thesetheoretical sale values had a lift of 2% to 11% over the observed sales,depending on the schedule. Since the algorithm provides multiplesolutions it is a call by the media planner to make a choice of the schedulethat best meets his requirement. In this case the solution we picked thesolution that largely mimicked the original media plan with a higherthreshold and lower peaks. This provided a lift of 6% over the observedsales value.
The NLP optimization used a fixed parameter (beta) over the entire period.The optimization will be improved drastically if one were to provide adynamic parameter for TV over time.